Thursday's cs.AI announcement day (2026-10-15) lists 108 new and 172 cross-lists (replacements skipped; listing total 280). A filter for agent systems, memory/context, computer-use / GUI / tools / skills / harnesses, multi-agent, persistence/identity, and local/open models keeps 33 papers. Personal and procedural memory share the lead (PraxisMemBench, multimodal memory, Cortex consolidation, transparent compaction, STRATA, XferAgent). Skills and harnesses follow with AtomSearch, CoRedSkill, marketplace attack-surface profiling, Doc2Harness, and L3M. On-device and small-model work (ixi-sLM, right-sizing, verified procedures) sits beside GUI recoverability, coding-agent report gaps, LAGO's growing options library, AgenticRV monitors, and know/act/hold self-awareness.
Research Papers
Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants
Maosheng Qin; Qi Zhang; Mingxuan Xia et al. arXiv: 2610.20022
PraxisMemBench evaluates harness-based persistent personal assistance where personal memory about the user and procedural memory about how to act must evolve together. Keeping the two types separately is insufficient: completion drops from 84.6% on single-type dependencies to 77.9% when tasks need both, and falls further when procedures must adapt to personal updates.
Key insight: Personal and procedural memory have to co-evolve: separate stores drop completion from 84.6% to 77.9% once tasks depend on both.
Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory
Minghao Guo; Qingyue Jiao; Xiujin Liu et al. arXiv: 2610.19366
A two-axis framework for multimodal agent memory separates how multimodal information is formed and represented from how memory management adapts with accumulated experience, targeting continuity across sessions and lifelong settings.
Key insight: Multimodal agent memory design splits into representation choices and adaptive management policies — they are separate levers.
Cortex: Consolidating Context into Parametric Memory
Jintian Zhang; Feng He; Xinnian Liang et al. arXiv: 2610.20399
Cortex consolidates context gains into mergeable low-rank or feed-forward updates via prospective queries, aiming to keep non-target capabilities. Across 4B–120B dense and MoE backbones it sits on the consolidation–retention Pareto frontier in most settings.
Key insight: Context gains can be internalized into mergeable parametric updates without forever-growing prompts, with measured retention under sequential edits.
Maximising Transparency in Context Compaction Improves Retention of Facts and Commitments for Lightweight Coding Agents
Felix Woolford arXiv: 2610.20116
transparent-compact is a deterministic compaction method for the open-source Pi coding agent that leaves user and assistant prose unchanged and explicitly marks where information was stripped, targeting lightweight and local models.
Key insight: For lightweight coding agents, transparent truncation with gap markers retains facts and commitments better than opaque model-written summaries.
STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents
Guanhua Zhu; Hanwen Li; Zhi Wei et al. arXiv: 2610.19386
STRATA organizes agent experience into a three-layer graph of atomic skills, workflows, and meta-frameworks with typed edges and self-maintenance mechanisms, aiming to stop negative transfer from flat embedding pools that copy opposite lessons.
Key insight: Structured, self-maintaining experience graphs reduce negative transfer that flat similarity-indexed memory pools invite.
XferAgent: Cross-Agent Experience Transfer for Self-Improving LLM Agents
Weihao Li; Jun Bai; Shijie Chen et al. arXiv: 2610.19581
XferAgent maintains local and shared experience graphs so heterogeneous agents can transfer compatible lessons without unreliable merges, selecting shared experiences per new task.
Key insight: Cross-agent experience transfer works when merges are reconciled on typed graphs rather than dumping raw traces across frameworks.
AtomSearch: Evidence-Grounded Tree Search for Agent Skill Optimization
Jia Tong; Senyou Deng; Min Peng et al. arXiv: 2610.19951
AtomSearch revises agent skills through contrastive atomic rule edits and a Skill-State Tree Search that keeps revisitable intermediate states. With deepseek-v4-flash it beats SkillOpt on three of four tasks, including a 9.68-point gain on one.
Key insight: Optimize skills one atomic rule at a time with search that retains useful mid-states, rather than rewriting entire skill files at once.
CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection
Qichen Zhao; Shengfang Zhai; Qingni Shen et al. arXiv: 2610.20049
CoRedSkill embeds attack operations across cooperating skills via AST rewriting so individual-package audits miss the composed intent, while each skill keeps its original utility. A 150-case benchmark demonstrates the gap.
Key insight: Skill composition is an attack surface: malicious workflows can hide across packages that look benign alone.
Beyond Maliciousness: Static and Runtime Attack-Surface Profiling of Agent Skills
Ratnesh Pandey; Japneet Singh; Harsh Akshit et al. arXiv: 2610.19264
Two-phase static and sandbox profiling of 6,309 marketplace skill listings finds shell-execution capability in about half, with substantial shares granting filesystem and network access — measuring the authority installed skills grant.
Key insight: Installing a marketplace skill often grants shell, filesystem, or network authority; profile the capability surface before trust.
Why Static Skill Inconsistencies Do Not Equate to Runtime Risks: A Behavioral Analysis of LLM Agent Execution
Fengrui Fan; Huahao Wang; Jingyi Fang et al. arXiv: 2610.18927
Behavioral analysis shows static inconsistencies in skill specifications often do not equal runtime risk, clarifying when static checkers over-flag versus miss real failures.
Key insight: Static skill inconsistencies are a weak proxy for runtime risk — pair linters with behavioral traces.
Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents
Yuyao Ge; Shenghua Liu; Peizhong Gao et al. arXiv: 2610.20344
Doc2Harness benchmarks whether models can turn long documents into working coding-agent harnesses over long horizons, measuring harness engineering rather than task solving inside a fixed harness.
Key insight: Long-horizon harness engineering from docs is its own capability, distinct from solving tasks inside an existing harness.
L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents
Carson Eisenach; Robert Joseph George; Vincent Quenneville-Belair et al. arXiv: 2610.19591
L3M provides a verifiable harness and runtime for agents that extend and improve themselves, gating self-modification behind checks so extensions do not silently break prior guarantees.
Key insight: Self-extending agents need a verifiable runtime gate before promoting their own harness or tool edits.
ixi-sLM: On-Device sLM for AI Call Agent Service
Han-Sang Lee; Kyuho Lee; Juneyoung Park et al. arXiv: 2610.19382
ixi-sLM is an on-device small language model for an AI call-agent service, keeping telephony-agent context local rather than shipping it to the cloud.
Key insight: On-device small models can host personal call agents without sending conversation context off-device.
When Is Smaller Greener? Right-Sizing Language Models for Agentic AI
Sajid Siraj; Maria Barbati arXiv: 2610.19345
When Is Smaller Greener? studies when smaller language models are greener and cheaper for agentic workloads, arguing for right-sizing rather than defaulting to frontier models for every step.
Key insight: Agentic workloads often get greener and cheaper by right-sizing the model per step instead of always calling a frontier model.
Compiling Expert Procedures into Verified Programs: When a Small Model Is Enough
Vishvesh Bhat; Asad Aali; Muhammad Ahmed Mohsin arXiv: 2610.19148
Compiles expert procedures into verified programs so small models can execute procedural steps reliably when verification is available.
Key insight: Verified programs can make small models enough for procedural agent steps that do not need open-ended generation.
Why Do Agent Teams Fail Across Boundaries? A Failure Taxonomy for Agent-to-Agent Systems with Private States
Sarath Shekkizhar; Prathyusha Jwalapuram; Zixuan Ke et al. arXiv: 2610.19342
A failure taxonomy for agent-to-agent systems with private boundaries catalogs why teams fail when crossing organizational or agent boundaries.
Key insight: Cross-boundary agent handoffs fail in distinctive ways — treat auth, schema, and private-state gaps as first-class failure classes.
Smart Alone, Blind Together: Why Accuracy Fails to CertifyTheory of Mind in Multi-Agent Teams
Will Beaumaster; Paul Schrater arXiv: 2610.19051
Shows that accuracy alone fails to certify Theory of Mind in multi-agent teams: agents that look competent solo still mis-model teammates.
Key insight: Solo accuracy does not certify teammate modeling; multi-agent evaluations need explicit Theory-of-Mind probes.
PFedMA: Who Should Join Next as Federated Multi-Agent Orchestration Evolves?
Shuai Li; Shengyuan Liu; Fan Qi et al. arXiv: 2610.19458
PFedMA studies who should join next as federated multi-agent orchestration evolves, separating membership selection from task policy.
Key insight: Federated multi-agent systems need an explicit membership policy as orchestration evolves, not only a task policy.
Task-State Governance: A Virtual Machine for Verifiable and Recoverable GUI Agent Execution
Wenkui Yang; Yiming Hao; Kai Zhou et al. arXiv: 2610.20100
Task-State Governance introduces a virtual machine for verifiable and recoverable GUI agent execution, so task state can be audited and rolled back.
Key insight: GUI agents need a recoverable task-state ledger, not only opaque screenshot loops.
Teaching Agents in their Own Style: Blackbox Teacher Distillation for Web Agents
Md Farhan Ishmam; Kenneth Marino arXiv: 2610.19834
Blackbox teacher distillation teaches web agents in their own action style, aligning student decisions without requiring teacher logits or matching tokenization.
Key insight: Frontier web-agent behavior can be distilled into students at the action level without white-box teacher access.
DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents
Tianle Chen; Pushkar Singh; Ketan Todi et al. arXiv: 2610.19376
DeLoop uses repeated actions in trajectories as self-supervision for long-horizon agents, turning loop patterns into training signal.
Key insight: Repeated actions in long-horizon trajectories are usable self-supervision, not only noise to suppress.
One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness
Artin Tajdini; Manpreet Kaur; Peng Qi arXiv: 2610.20576
Simple cross-interface consistency training improves LLM agent robustness when the same task is exposed through different UIs or APIs.
Key insight: Training for consistency across interfaces improves agent robustness when the same task arrives via different channels.
What Coding Agents Leave Unsaid: Measuring Task-State Gaps Between Execution and Reports
Guoyang Xu; Rui Zhang; Xiuyu Wu et al. arXiv: 2610.20003
Measures task-state gaps between what coding agents actually execute and what they report, revealing silent omissions in status updates.
Key insight: Coding agents often omit executed state from their reports — status text is not a complete task ledger.
Co-Evolving Acting and Forecasting for Language Agents
Haofei Yu; Yixing Chen; Yining Zhao et al. arXiv: 2610.20597
Adds multi-step action forecasting on successful on-policy trajectories so the policy predicts upcoming actions from each history, denser than terminal reward alone.
Key insight: Supervising multi-step action forecasts on successful trajectories densifies long-horizon agent learning beyond terminal rewards.
LAGO: Language Agents with Reusable Growing Options
Yuran Wu; Mengkang Hu; Qionglin Qiu et al. arXiv: 2610.19843
LAGO treats specialized language agents as reusable options in a growing library: a meta-agent retrieves or spawns options and admits successful ones without weight updates. Frozen Qwen3.5-35B-A3B reaches 58.6% on SWE-bench Verified; on a matched 150-task subset it beats mini-SWE-agent and Terminus 2.
Key insight: A growing library of reusable agent options (no weight updates) reaches 58.6% on SWE-bench Verified with a frozen 35B-A3B model and beats strong single-loop baselines under matched eval.
One Monitor Does Not Fit All: Formalism-Agnostic Runtime Verification for Agentic AI
Angelo Ferrando arXiv: 2610.19818
AgenticRV separates execution boundaries from monitoring formalisms: adapters emit canonical events, monitors receive projections, and a decision policy can block before effects. Distinguishes rejected attempts from completed effects; Python support includes LangGraph.
Key insight: Runtime verification for agents should be able to block before tool effects and must not confuse a rejected attempt with a completed violation.
AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety
Chenxu Wang; Xiaoyu Wen; Han Qi et al. arXiv: 2610.19422
AgentSafeEvo co-evolves an attacker that builds adversarial agent tasks with a defender that must complete tools safely, spanning diverse tasks and attack types with learnability-guided weighting.
Key insight: Adversarial co-evolution over tool-using agent tasks strengthens safety beyond static prompt-injection suites.
Agentic Self-Awareness: Aligning Knowing, Acting, and Holding
Qinglin Chen; Hiroaki Hayashi; Jiaxin Zhang et al. arXiv: 2610.19226
Agentic self-awareness aligns feasibility judgments (know), warranted behavior (act), and reconsideration (hold). On Terminal-Bench protocols, blocker recognition reaches 84.4% but strict self-awareness only 40.6%; harness search without weight updates helps.
Key insight: Recognizing a blocker is not enough: strict know/act/hold alignment is much rarer (40.6% vs 84.4% recognition) and harness search can improve it without retraining.
SERO: Learning from Executable Oracle Continuations for Long-Horizon Agentic Reinforcement Learning
Dingyi Rong; Yue Shi; Chaofan Ma et al. arXiv: 2610.19397
SERO restores a state the current policy reached and lets a training-only oracle continue in the real environment, making oracle futures available as training trajectories for long-horizon agentic RL.
Key insight: When on-policy rollouts never reach good futures, oracle continuations from reachable states supply supervision that reweighting failures cannot.
Structure Before Intelligence: Gating What Agentic Scientific Workflows Release
Ashish Acharya; Aman Al Amin; Hong Qin arXiv: 2610.19507
A harness around a scientific workflow DAG decides which outputs may be released: mandatory checks, branch containment, versioned recompute, and a risk model. Language-model agents propose fixes inside the harness but cannot override release gates.
Key insight: Agentic workflows need a release harness: agents may propose, but only checked, contained outputs should ship.
ActionOPD: Distilling Decisions, Not Tokens
Huining Yuan; Xiangmin Yi; Zelai Xu et al. arXiv: 2610.19820
ActionOPD distills agent decisions at semantic-action granularity on student-encountered states, working with black-box teachers that need not share tokenization or expose token probabilities.
Key insight: Distill capable agents into smaller students by aligning actions on student states, not tokens — black-box teachers are enough.
The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery
Robert Tjarko Lange; Luca Grillotti; Yutaro Yamada et al. arXiv: 2610.18933
The Agentic Scientist orchestrates coding agents through drafting, ideation, planning, sandboxed experiments, baselines, and reporting, keeping humans in the loop to offset limited ideation radius of fully autonomous research agents.
Key insight: Human-orchestrated coding-agent pipelines beat fully autonomous research agents when ideation breadth matters.
$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
Rishitosh Kumar Singh; Amir Saeidi; Shamanthak Hegde et al. arXiv: 2610.19560
τ-Emotion-Bench evaluates tool-calling agents under emotion-conditioned users across ten domains while preserving task objectives; Tracer generates grounded trajectories and lets dissatisfied users terminate.
Key insight: Tool-calling agents fail from user-interaction fragility under emotional variation, not only from tool-use errors.